Skip to content

AI data market

Synthetic data business model comparison: who cheap generation helps

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Synthetic data is five businesses, not one: generation tooling, generation services, licensed seeds, validators and finished synthetic datasets. Cheaper generation presses hardest on the models that sell reproducible output, such as tooling and generic datasets, while seeds and validators hold value only where they supply real-world grounding or a check a generator cannot run on itself. A matched-budget contrast ladder prices each claim.

Value drivers

Synthetic data business

This page covers synthetic data business model comparison: who cheap generation helps.Illustrative

Key takeaways

  • Synthetic data companies sell one of five things, and cheaper generation helps some of them and hurts others.
  • Tooling and generic finished datasets face the most pressure: permissive SDKs and large open releases set their reference price.
  • Licensed seeds and checks against real outcomes can hold value because a generator cannot supply grounding or verify itself.
  • Substitution is real for checkable tasks, style work and privacy-constrained bias testing, and weak where value depends on recorded outcomes.
  • A six-arm matched-budget ladder prices each model through one contrast; ask any company which contrast it has run.

Is synthetic data one business or five?

Five. Companies selling synthetic data sell generation tooling, generation services, licensed seed data, validation, or finished synthetic datasets, and those five respond in opposite directions to the same trend. Treating them as one category is how an investor ends up paying a scarcity multiple for a business whose output anyone can now reproduce.

The common view runs in a straight line: synthetic data gets cheaper and better, so it replaces licensed real data, so real-data businesses shrink. The map in this article argues something narrower and easier to test. As generation cost falls, the models that sell generated output face price pressure from free code and open releases, while the models that sell what a generator cannot make for itself, real-world grounding and trustworthy checks, can hold value. Whether they do is an empirical question, and the experiment below is designed to answer it for any one company.

The thesis fails if a generator with only public seeds and no outside checks matches the results of licensed seeds plus validation at the same budget, on real-world tests none of them touched. That outcome would mean the inputs-and-checks layer is not scarce after all.

What each model sells, and the public evidence on its price

Each model sells a different thing, and public evidence already shows free or near-free substitutes for several of them. Tooling and generic finished datasets show the clearest pressure; seeds and validation show leverage that depends on what the seed or the check contains.

Generation tooling is software that fits a generator to data and produces records. MOSTLY AI publishes its Synthetic Data SDK under the Apache 2.0 license, and its repository describes a local mode that trains and generates on the user's own compute. DataCebo takes a different route with SDV, licensed under the Business Source License 1.1 with a grant that allows use except for a Synthetic Data Service offered to third parties, each release converting to the MIT license four years later. Consolidation points the same way: SAS announced in November 2024 that it was acquiring the principal software assets of Hazy to extend its synthetic data product, and Gretel's site now says it is part of NVIDIA. Standalone generation software is becoming a feature of larger platforms.

Generation services produce data to a buyer's specification. Their substitute is doing it yourself: Stanford's Alpaca team reported in 2023 that it generated 52K instruction-following examples with an OpenAI model for less than $500, building on the Self-Instruct pipeline. Licensed seeds are real records used to condition or anchor generation; the same Alpaca repository starts from a seed file of 175 tasks, and Hugging Face built Cosmopedia's prompts from seed samples meant to cover the knowledge in web datasets such as RefinedWeb and RedPajama.

Validators decide which generated examples to keep. Where an answer can be checked mechanically, the check is code: the Tülu 3 report (2024) trains its final stage on verifiable rewards instead of a reward model, for skills such as mathematics and precise instruction following. Finished synthetic datasets are products in their own right, and open releases set their reference price. Cosmopedia, which Hugging Face's repository (last updated November 2024) describes as over 30 million files and 25 billion tokens generated by Mixtral-8x7B-Instruct in more than 10,000 H100 GPU hours, was released openly, while Alpaca's 52K examples carry a CC BY NC 4.0 license that allows only non-commercial use.

Synthetic data business model comparison: the substitution matrix

The matrix rates each model on two questions an investor can check: how hard free or cheap substitutes press on its price, and how much it needs real-world data to deliver what buyers pay for. Validators split into two rows because a check code can run and a check against real outcomes behave like different businesses.

Ratings are editorial judgments drawn from the dated evidence in the third column, written on 2026-10-27; they are a starting hypothesis for diligence, not a measurement of any company. The rows:

Synthetic data business model comparison: the substitution matrix
ModelWhat the buyer pays forCheapest substitute, with evidenceNeed for real-world dataSubstitute pressure
Generation toolingSoftware to fit generators and produce recordsOpen SDKs such as MOSTLY AI's Apache 2.0 release; SDV source-available with a service carve-outLow: the customer brings its own recordsHigh
Generation servicesGenerated data built to a specificationGenerating in-house through model APIs; Alpaca's 52K examples for under $500 (2023)Low, unless the vendor holds domain seedsHigh for generic specifications
Licensed seedsReal records that condition or anchor generationPublic text and synthetic seeds; WRAP rephrases public web pagesHigh: the records are the productMedium
Validators: checks code can runAccept or reject decisions on output with checkable answersVerifiable rewards and rule checks, as in Tülu 3 (2024)LowHigh
Validators: checks against real outcomesJudgments that compare output with what actually happenedModel judges, where their agreement with outcomes holds upHigh: reference outcomes come from recordsLow
Finished synthetic datasetsA ready-made dataset under a licenseOpen releases such as Cosmopedia's 25 billion tokens (2024)Medium: seeds sit upstreamHigh for generic corpora
A three-by-three map of substitute pressure against need for real-world data: real-outcome checks and licensed seeds sit in the high-need, low-to-medium-pressure zone, while tooling, services, code-run checks and finished datasets sit under high pressure, with an arrow showing value moving toward inputs and checks.
Six positions for five synthetic data business models, placed from the substitution matrix ratings. Editorial ratings, not measured market data.

Why cheaper generation pushes value toward inputs and checks

Because the parts of the stack that can be copied get copied, and their price falls toward the cost of compute. What cannot be copied by running a generator longer is knowledge the generator does not have and confidence that an output is right, so value collects in the seed and validation rows.

The license choices of tooling vendors already reflect this. SDV's grant permits broad use but carves out offering it as a Synthetic Data Service, which protects the service revenue a free tool would otherwise hand to competitors. MOSTLY AI chose the opposite trade, a fully permissive license, which reads as treating the code as a way to reach users rather than as the thing being sold. Either way, the license protects or gives away output capability; neither can make an input scarcer.

Research on rephrasing shows where the knowledge sits. The WRAP paper (Maini and colleagues, 2024) had an instruction-tuned model rewrite web documents in styles such as Wikipedia or question-and-answer, trained on real and rephrased text together, and reported roughly three times faster pre-training on the C4 dataset. Its authors note that a small open model can do the rewriting because the method does not rely on the model as a knowledge bank. The real documents carry the facts; the generator carries the style. A seed business sells the first of those, and the second is getting cheaper every year.

Where synthetic data does replace licensed records

In three places at least: tasks with checkable answers, format and style work, and settings where law or privacy pushes toward generated data. An honest map gives these their full weight, because they are where the bear case for licensed data is strongest.

Checkable tasks come first. Tülu 3's final stage scores answers with a program instead of a reward model for the skills it targets, so for mathematics and constrained instruction following the check itself is code rather than purchased judgment. Style and format follow: WRAP's speedup came from restyling documents that were already in the training set. Regulation is the third. As adopted in 2024, Article 10(5) of the EU AI Act lets providers of high-risk AI systems process special categories of personal data for bias detection only where that cannot be done effectively with other data, 'including synthetic or anonymised data'. That scope is limited to high-risk systems, and EUR-Lex lists a consolidated version dated 27 July 2026, so check the current text; the direction is still clear, since law can make synthetic data the first resort.

None of these removes the need for records that describe how a business actually ran. A verifiable reward needs a checkable answer, and most operational decisions, such as whether a claim should have been paid or a dispatch was right, are judged by outcomes stored in company systems. Substitution is real at the checkable, stylistic and privacy-constrained edges, and weak at the outcome-dependent center.

A contrast ladder: one matched-budget experiment that prices each model

A single matched-budget experiment can price all five models at once if its arms add one layer at a time. Each adjacent pair isolates one business model's contribution, so a company claiming value in one row can be asked for the contrast that row implies.

Hold the budget fixed across arms and count everything in it: data acquisition or license fees, generation compute, validation labor and any vendor charge, at quoted prices. Use the same base model, training recipe and number of training runs per arm, so run-to-run variance is visible. Score every arm on an untouched real-world holdout drawn from companies and a time window no arm used for seeds, generation or tuning, and fix the contrasts before any results arrive.

SourceX has not run this experiment and reports no results from it; the design is published so investors can request it from any company in the space. The ladder:

A contrast ladder: one matched-budget experiment that prices each model
ArmWhat it addsContrastWhat the contrast prices
A. Open tools, public seeds, no filteringThe free baselineA on its ownHow far free tooling goes
B. Arm A with licensed real seedsReal-world groundingB minus ALicensed seeds
C. Arm B filtered by a validator to the same final countA check on every kept exampleC minus BValidators
D. A vendor's service or finished dataset at the same budgetThe vendor's productD minus CServices and finished datasets
E. Licensed real records onlyReal data in place of generated dataE compared with CWhether synthetic substitutes at this budget
F. Half the budget as E, half as CA mixtureF compared with C and EWhether real and synthetic complement each other

Five ways the map could misprice a company

The map could be wrong in at least five ways, and each one points to a different company being mispriced. Investors should treat these as live scenarios, not footnotes.

Seed demand may be tiny. Alpaca's 2023 pipeline produced 52K examples from 175 seed tasks, so a seed seller may earn a one-time fee per domain rather than revenue that grows with a buyer's output. That favors businesses holding rare records many buyers need, and penalizes those holding common ones.

Outcome checks may automate. If model judges become reliable on business decisions, the low-pressure validator row moves right on the map, and real-outcome data becomes a calibration set bought once rather than a recurring input.

Terms travel downstream and cut both ways. Alpaca's non-commercial license shows how an upstream term can block resale of a finished dataset; the same mechanism can protect a seed owner whose license controls what generated output may be used for, or strand a dataset vendor whose generator terms change.

Platforms may absorb whole rows. The SAS and NVIDIA moves suggest generation tooling increasingly sells as a feature, which narrows exits for standalone tooling companies to a small set of acquirers.

The experiment may find nothing at the budget tested. A null seed effect in one domain at one budget does not prove seeds worthless elsewhere, but it does mean a company's claim is unproven, and the burden of evidence should stay with the company.

What to ask each kind of company, and what to ask SourceX

Ask each company for the contrast that its row of the matrix implies, plus the license terms that decide who keeps the value. A company that cannot name its row, or claims value in several rows without evidence for each, is telling you something.

On this map, SourceX sits in the rows that depend on real-world data: it helps companies license selected business records to AI developers and runs the sourcing, rights review, contracting, preparation, delivery and payment, with datasets sourced when a buyer asks rather than held in stock. The same questions apply to it. Its commercial metrics are not public; SourceX shares them with investors directly. It is raising under Rule 506(c), so only verified accredited investors can take part, and anyone evaluating the company can ask for materials through the investors page. Questions by row:

Checklist · 7 items

  1. Tooling: what share of revenue depends on features that open SDKs lack, and what happens to pricing when a platform bundles generation?
  2. Services: what does the company supply that a buyer generating in-house through an API cannot, and is it seeds, checks or labor?
  3. Seeds: which uses do licenses permit (conditioning generation, training on outputs, keeping outputs after the term), and how is each priced?
  4. Code-run validators: which checks could a buyer's own engineers write, and what is left once they do?
  5. Real-outcome validators: where do the reference outcomes come from, who has rights to them, and how often must they be refreshed?
  6. Finished datasets: which generator and seed terms govern resale, and what does the product offer that an open release does not?
  7. Any company, SourceX included: which arm comparison from the ladder has it run, on what holdout, and will it share the design and the result?

0 of 7 ticked. Tick each item as you go.

How the map was built and what it leaves out

The matrix, the map and the contrast ladder were written for this article on 2026-10-27 as editorial tools. Evidence comes from the companies' own repositories and announcements (MOSTLY AI, DataCebo's SDV license, SAS, Gretel), the Stanford Alpaca and Hugging Face Cosmopedia repositories, the WRAP and Tülu 3 papers, and the EU AI Act as published on EUR-Lex; each figure in the text names its source and date.

Not measured: market sizes, company revenues, margins, growth or customer counts for any company named, and no company is ranked. Ratings in the matrix are judgments a reader can contest row by row, and the map's positions come only from those ratings. The ladder is a design with no results. Company descriptions rely on public statements and license texts that can change; recheck them before relying on any row.

Frequently asked questions

What are the main synthetic data business models?

Five recur: generation tooling (software), generation services (data made to a specification), licensed seed data (real records that condition generation), validation (checks that accept or reject outputs) and finished synthetic datasets. They face very different pressure as generation gets cheaper, so they should be valued separately.

Does synthetic data make licensed real data worthless?

Not on current evidence. Synthetic data substitutes well for tasks with checkable answers, for style and format work and where privacy rules favor it. It substitutes poorly where the value lies in facts or outcomes only real records contain, and real data often becomes the seed or the test for generated data.

Why is synthetic data tooling under price pressure?

Capable generation software is now available under permissive licenses, and large platforms are absorbing standalone vendors, as SAS did with Hazy's software assets in 2024. When buyers can run comparable tooling on their own compute, tooling revenue depends on features and support rather than the generation itself.

How can an investor test a synthetic data company's claims?

Ask for a matched-budget comparison in which arms add one layer at a time (free tools, licensed seeds, a validator, the vendor's product, real data only, a mixture), scored on a real-world holdout none of the arms touched. The difference between adjacent arms prices the layer the company sells.

Where does SourceX fit on the map?

In the rows that depend on real-world data: licensed records used as seeds, training data or reference outcomes. Its commercial metrics are not public; SourceX shares company metrics with investors directly, and investor materials are requested through sourcex.si/investors.

Sources

  • MOSTLY AI publishes its Synthetic Data SDK under the Apache 2.0 license, with a local mode that trains generators and generates data on the user's own compute. Source
  • DataCebo licenses SDV under the Business Source License 1.1, granting use except for a Synthetic Data Service offered to third parties, with each release changing to the MIT license four years after release. Source
  • SAS announced in November 2024 that it was acquiring the principal software assets of Hazy, a synthetic data company, to extend SAS Data Maker. Source
  • Gretel's website says Gretel is now part of NVIDIA. Source
  • The Stanford Alpaca repository reports 52K instruction-following examples generated with text-davinci-003 for less than $500, built on the Self-Instruct pipeline from a 175-task seed file, with the data licensed CC BY NC 4.0 for non-commercial use. Source
  • Hugging Face's Cosmopedia repository describes over 30 million files and 25 billion tokens of synthetic text generated by Mixtral-8x7B-Instruct in more than 10,000 H100 GPU hours, with prompts built from seed data covering web datasets. Source
  • The WRAP paper (2024) rephrases web documents in styles such as Wikipedia or question-answer, trains on real and rephrased text, and reports roughly 3x faster pre-training on C4 using a small open rephrasing model. Source
  • The Tülu 3 report (2024) describes reinforcement learning with verifiable rewards, trained on verifiable rewards instead of a reward model, for skills such as mathematics and precise instruction following. Source
  • Article 10(5)(a) of the EU AI Act, as adopted for high-risk AI systems, allows special categories of personal data for bias detection only where other data, including synthetic or anonymised data, cannot do the job effectively. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies